首页 > AI前沿 > Introducing Lockdown Mode and Elevated Risk labels in ChatGPT

Introducing Lockdown Mode and Elevated Risk labels in ChatGPT

OpenAI 2026-02-13 18:00 1 阅读 查看原文

Introducing Lockdown Mode and Elevated Risk Labels in ChatGPT

We are introducing Lockdown Mode and Elevated Risk labels in ChatGPT to help organizations defend against prompt injection and AI-driven data exfiltration.

These new security controls are designed to give security teams greater visibility and enforcement power over how ChatGPT interacts with external data and tools, particularly in high-stakes enterprise environments.

Lockdown Mode: Restricting Unnecessary Data Access

Lockdown Mode is a new configuration that allows administrators to strictly limit ChatGPT’s access to external resources, such as browsing, plugins, and custom actions, unless explicitly permitted by policy.

When Lockdown Mode is enabled, the model will:

  • Block all outbound network requests by default, including web browsing and API calls.
  • Prevent the use of third-party plugins or tools that are not on an approved allowlist.
  • Disable file uploads and downloads that could be used for data exfiltration.
  • Require explicit user or admin approval for any privileged action.

This mode is particularly useful for organizations handling sensitive data, where even accidental data leakage could have serious compliance or security implications.

Elevated Risk Labels: Real-Time Threat Detection

Elevated Risk labels provide real-time, contextual warnings when ChatGPT detects a potential prompt injection attempt or an unusual pattern that could indicate data exfiltration.

These labels appear directly in the conversation interface and in the admin audit logs, allowing security teams to:

  • Identify suspicious prompts that attempt to override system instructions.
  • Flag conversations where the model is asked to output hidden system prompts or internal data.
  • Detect repeated requests to send data to external URLs or email addresses.
  • Receive a severity rating (Low, Medium, High, Critical) for each flagged event.

Important: Elevated Risk labels are not a replacement for dedicated security monitoring. They are designed to complement existing security information and event management (SIEM) systems by providing AI-specific threat intelligence.

How These Features Work Together

When both features are enabled, organizations can create a defense-in-depth strategy against AI-specific attacks. Lockdown Mode reduces the attack surface by limiting what the model can access, while Elevated Risk labels provide early warning signs of malicious or anomalous behavior.

For example, if an attacker attempts to use a prompt injection to force ChatGPT to send internal data to a malicious server, Lockdown Mode would block the outbound request, and the Elevated Risk label would immediately flag the conversation for review.

Availability and Configuration

Both features are available today for ChatGPT Enterprise and Education customers. Administrators can enable them via the admin console under Security Controls.

We recommend starting with a pilot group to fine-tune policies before rolling out organization-wide. For detailed configuration guidance, please refer to our help center documentation.

# Example admin policy snippet (YAML)
security:
  lockdown_mode: true
  allowed_tools:
    - "internal-db-query"
    - "approved-calculator"
  elevated_risk_labels:
    enabled: true
    severity_threshold: "medium"
    notify_security_team: true

We will continue to enhance these capabilities based on customer feedback and emerging threat intelligence. Stay tuned for further updates.